AI Value Attribution Framework

The AI Value Attribution Framework is Crescent Capital Advisors' five-level model for verifying what AI investment produced: Cost, Adoption, Productivity, Business Outcomes, Enterprise Value. Most portfolio companies can evidence the first two levels and report as though they had evidenced the last two. The framework names that gap and supplies the four questions that close it before the next initiative gets funded.

Why a Framework and Not a Dashboard

A dashboard answers the questions someone thought to instrument. Most AI dashboards were specified during the pilot phase, when the interesting questions were about usage: who logged in, how many prompts ran, which agents fired. Those panels keep reporting the same things long after the budget conversation has moved on to what the spend bought.

The problem sits a level above the tooling. A company can hold every number its dashboard produces and still connect none of them to a line in the P&L, because the baseline was never captured, no one owned the result, or the outcome got defined after the fact to fit whatever happened to move.

So the framework holds five levels in a fixed order, each with its own question, its own metrics, and a plain statement of where it stops. A leadership team can walk the ladder in a morning and find the rung where its evidence runs out. That rung is the claim it can defend in a board meeting, and everything above it is narrative.

The Five Levels of AI Measurement

Each level answers a different question and satisfies a different audience. The status column is the pattern CCA sees across mid-market portfolio companies: how many can produce evidence at that level today without building new measurement.

Level	Name	Executive question	Status across most companies
1	Cost	How much are we spending?	Most can measure
2	Adoption	Are people using it?	Most can measure
3	Productivity	Are teams working faster?	Some can measure
4	Business Outcomes	Did business results improve?	Very few can measure
5	Enterprise Value	Did the economics of the business change?	Almost none can measure consistently

Levels 1 and 2 describe activity. Levels 4 and 5 describe economics. Level 3 is the crossing point, and it's where most AI value stories stall out: velocity improved, everyone agrees it improved, and no one carried the improvement into a financial statement.

Read the status column against the audience column. Boards, lenders, and buyers ask Level 4 and Level 5 questions. The measurement most companies built answers Levels 1 and 2. That mismatch is the attribution gap.

Level 1: Cost

Question: How much are we spending?

Typical metrics: AI software licenses, API spend, model costs, infrastructure, token consumption.

What it tells you: whether spending is rising, and how fast.

Where it stops: cost says nothing about value, and it is routinely understated. See the AI Cost Optimization crosswalk below for why the stated number is usually the wrong denominator.

Level 2: Adoption

Question: Are people using it?

Typical metrics: active users, prompt volume, agent executions, session frequency, seats assigned versus seats used.

What it tells you: whether the deployment reached the people it was bought for.

Where it stops: usage counts the input, not the result. A team can run 40,000 prompts a quarter and close the same number of tickets it closed before.

Level 3: Productivity

Question: Are teams working faster?

Typical metrics: development cycle time, support handling time, content production velocity, time to completion, rework rate.

What it tells you: whether AI changed how the work gets done. This is the first level with a real signal in it, and it's where most programs plant their flag.

Where it stops: hours saved become money only when the hours get redeployed or removed. A 30% handling-time improvement with flat headcount and flat volume is a claim about capacity, and capacity converts to economics only when someone decides what to do with it.

Level 4: Business Outcomes

Question: Did business results improve?

Typical metrics: customer onboarding time, support resolution rates, renewal and gross retention, customer satisfaction, sales productivity, lead conversion.

What it tells you: whether AI influenced the operating results a CEO already reports monthly.

Where it stops: attribution gets hard here, because everything else in the business also moved. Pricing changed, a competitor stumbled, the team hired. A Level 4 number holds up only when a baseline was captured before the deployment and the confounders are named.

Level 5: Enterprise Value

Question: Did the economics of the business change?

Typical metrics: EBITDA impact, revenue growth, gross margin, net revenue retention, operating leverage, and the AI narrative a buyer will price.

What it tells you: whether AI moved the number the sponsor is underwriting.

Where it stops: Level 5 inherits every weakness below it. A Level 5 claim built on an unverified Level 4 outcome, on an understated Level 1 cost, is arithmetic performed on two guesses.

Why the Gap Exists

Most AI programs start as experiments, and experiments are measured on learning and usage. That's the correct instrumentation for the first two quarters. Then spending grows, the CFO starts asking about run rate, and the question changes from what AI can do to what AI did.

The measurement system that would answer the second question had to be built before the deployment, because it needs a baseline. By the time anyone asks, the pre-AI number is gone.

Three Failure Modes

Adoption Theater. Leadership reports user counts, prompt volume, and agent executions with real pride. Nobody in the room can state the business impact. Activity has been accepted as a proxy for value, and the proxy holds right up until a board member asks for the second-order number.

Productivity Without Economics. Teams genuinely got faster. Cycle time fell, the engineers can feel it, the survey scores back it up. No one connected the improvement to a financial line, so the gain sits in a slide rather than in the P&L, and the finance team discounts it to zero when the budget gets built.

Cost Visibility Without Value Visibility. Finance can account for every dollar spent on AI. Nobody can account for a dollar earned, saved, or protected. This is the failure mode that kills programs, because a cost with no matching benefit is the easiest line to cut in a tight quarter, and the cut lands on the initiatives that were working alongside the ones that weren't.

The Four Pre-Approval Questions

Every AI initiative answers these before it gets funded. They take a meeting. Skipping them costs a year of evidence that can't be reconstructed afterward.

1. What business outcome are we trying to improve? Named at Level 4 or Level 5, in the language the company already uses on its monthly reporting.
2. How will we measure it? The specific metric, the system it comes from, and the reporting cadence.
3. What baseline are we comparing against? Captured before deployment, with the known confounders written down while everyone still remembers them.
4. Who owns the result? One accountable executive, named in the approval. A vendor can't hold this one, and an AI working group holds it collectively, which is the same as nobody.

The fourth question fails most often, and it fails quietly. An initiative with a shared owner produces a shared explanation when the number doesn't move.

Self-Assessment Rubric

Attribution maturity is a ladder rather than a score. Your defensible level is the highest level where you hold evidence and every level beneath it also holds. A gap anywhere below caps the claim above it, which is why most companies that believe they're operating at Level 4 are defensible at Level 2.

Level	You can claim it when	What caps the claim
1	Total AI cost includes labor, waste, and risk, and rolls up to a named feature owner	An invoice-only figure that counts vendor spend and stops there
2	Usage split by team and workflow, with seats assigned measured against seats used	A license count reported as an adoption number
3	Cycle time or handling time measured against a pre-deployment baseline from the same system	A survey asking people how much time they think they saved
4	An operating metric moved, the baseline predates the deployment, and the confounders in the period are documented	A result claimed in a quarter when pricing, staffing, or product also changed
5	The Level 4 outcome is priced into EBITDA or margin, against a Level 1 cost that includes labor and waste	An EBITDA number derived from hours saved that were never redeployed or removed

Run the ladder per initiative rather than per company. A single portfolio company will hold a Level 4 claim on one workflow and a Level 2 claim on five others, and the aggregate number it reports to the board averages them into something no one can defend.

An Illustrative Attribution Walk

The figures below are modeled to show how the levels interact. They aren't results from a named CCA engagement.

A portfolio company deploys AI across customer support. Finance reports AI spend of $270K a year, covering licenses, API, and infrastructure. Support handling time falls from 14 minutes to 9.8 minutes across 22,000 tickets a quarter. The team reports a 30% productivity gain and, separately, credits AI with two points of gross retention on $14M of ARR. Combined, that's a $592K benefit on $270K of spend, and the deck says 2.2x.

Walk the ladder and the number changes.

Level 1. The $270K covers the first three cost layers, which carry 45% of true AI spend under the AI Cost Optimization model. Labor at 40%, waste at 10%, and risk at 5% never entered the figure. True annual cost is closer to $600K.

Level 3. The 4.2 minutes saved per ticket across 88,000 tickets a year is 6,160 hours, which is real and measured against a baseline the ticketing system already held. At 1,700 productive hours per agent, that's 3.6 FTE of capacity.

Level 4. Volume grew 19% and the team didn't hire. Four agents that the old ratio would have required were never posted, worth $312K fully loaded. That claim survives, because the baseline predates the deployment and the ratio is documented. The retention claim doesn't survive: a pricing change shipped in the same quarter, and no one wrote down which accounts were touched by which. It fails question 3.

Level 5. Verified benefit of $312K against true cost of $600K puts the program at roughly 52 cents on the dollar in year one, against a reported 2.2x. The gap between those two numbers is the entire argument for doing this work.

That finding is useful rather than fatal. It hands the operating partner two live moves: run the Cost Optimization pass on the $600K, which returns 20 to 34% in year one, and narrow the deployment to the workflows carrying the verified benefit. Do both and the position turns inside a quarter, on a program that would otherwise have been cut on the strength of a number nobody trusted.

Board Questions a Leadership Team Should Be Able to Answer

Seven questions. A team that can answer all seven with evidence is operating at Level 4 or better.

• Where are we spending on AI, including labor and waste?
• Which initiatives are creating measurable value?
• Which ones aren't?
• Which investments should be expanded?
• Which should be stopped?
• What economic outcomes have been realized to date?
• How is AI affecting operating leverage?

The two hardest questions on that list are the third and the fifth. A program with no answer to them has no stopping mechanism, which means its cost grows and its scrutiny grows with it.

Crosswalk: The Three Economic Lenses

CCA runs three AI economics frameworks. Each answers a different executive question, and the vocabulary is locked per CCA decision D-076 so that no framework borrows another's verb.

Question the executive is asking	Framework that owns it	Verb
Where should the next AI dollar go?	AI Value Creation	Allocate
Which dollars are already being wasted?	AI Cost Optimization	Recover
Did the investment produce anything?	AI Value Attribution	Verify

AI Value Creation identifies where AI capital should be deployed across the six AI Investment Categories. This framework verifies whether the deployment produced anything. Run in sequence, capital allocation is followed by outcome verification, and the verification result feeds the next allocation cycle.

AI Cost Optimization works the same Level 1 number from the opposite direction. Attribution treats cost as the denominator it measures value against; Cost Optimization treats that denominator as the thing to shrink. Its 6 Cost Layers explain why Level 1 is under-measured almost everywhere: people run 40% of AI spend and waste runs 10%, and neither appears in a stated AI cost figure built from vendor invoices. A company measuring Level 1 badly will overstate Levels 4 and 5 by the size of the missing 55%.

CLEAR™ runs attribution as a workstream inside Leverage and Accelerate. Leverage deploys AI against EBITDA; attribution verifies that the EBITDA showed up and names the initiatives that produced it. Accelerate carries the same test for revenue-side deployments, where the confounders are heavier and the baseline discipline matters more.

PRISM™ gates the whole exercise through the Strategic Data Assets dimension, scored in the AI Readiness Index. Levels 4 and 5 need outcome data that's joined, historical, and trusted. A company scoring low on data foundations can't measure past Level 2 regardless of intent, and the honest first move there is fixing the data rather than commissioning an attribution study that will return noise.

Frequently Asked Questions

What is AI Value Attribution? The practice of connecting AI spending to measurable business outcomes. It runs on a five-level hierarchy (Cost, Adoption, Productivity, Business Outcomes, Enterprise Value) and its output is a defensible statement of which AI initiatives produced economic results and which didn't.

How is this different from measuring AI ROI? An ROI figure is a single output. This framework is the method that makes the output defensible: it names the level your evidence reaches, forces a baseline before deployment, and requires an owner on every claim. Most stated AI ROI numbers are Level 3 productivity gains presented in Level 5 language.

Why do most organizations stall at Level 2? Adoption metrics arrive free with the tooling, and outcome metrics have to be designed before the deployment. By the time leadership asks the outcome question, the pre-deployment baseline no longer exists.

Can we start attribution on a program that's already running? Yes, with a smaller claim. Historical baselines can sometimes be reconstructed from systems that were logging anyway, such as ticketing, CRM, or version control. Anything that depended on a survey or a manual process usually can't be recovered, so the practical move is to apply the four questions to everything not yet approved and rebuild what's reconstructable behind it.

Who owns AI value attribution inside a portfolio company? The CFO owns the measurement standard and the reporting; a named business owner owns each initiative's result. Attribution held inside an AI center of excellence tends to grade its own homework.

What does this cost in effort? The four pre-approval questions add a meeting per initiative. Building the measurement for a first cohort of initiatives is a few weeks of finance and data engineering time, most of it spent defining metrics rather than building pipelines.

How does this connect to exit value? A buyer discounts an AI story it can't verify, and a diligence team will ask for the baseline. Attribution evidence converts an AI narrative into an underwritable line, which is what the CLEAR™ Realize phase is protecting.

Does this apply to a company still running AI pilots? That's the cheapest time to apply it. Baselines captured while the estate is small cost almost nothing, and they're the asset that makes every later claim defensible.

Next Step

Bring the AI spend, the initiative list, and whatever measurement exists today. CCA walks each initiative up the five levels, marks the rung where the evidence runs out, and returns the defensible number alongside the four questions applied to whatever's next in the queue. Book a call to run it against a specific portfolio company.

v1.0 · Updated August 2026